ZJIT: Big projects 2026

[tekknolagi]

• Stack map for local variables
• Come up with a unified theory of local variables
• Come up with a unified calling convention
  • Delete or “see through” array allocations for
    • *rest parameter passing
    • caller *splat argument
• Design and implement compiler<->interpreter state machine for function lifecycle
• Speed up the compiler
  • Reduce memory usage of Insn/lir::Insn ZJIT: Shrink Insn and InsnId #986
  • Minimize SSA: ZJIT: Add Braun SSA optimization pass ruby/ruby#17311
  • Don’t use full backend for generating/rewriting/patching stubs ZJIT: Stop running trampolines and stubs through full LIR backend #963
• Extend optimization passes
  • Make load-store et al whole-CFG instead of block-local
  • Make canonicalization whole-CFG instead of block-local ZJIT: Dominator-scoped canonicalize for cross-block redundant guard elimination #983
• Improve inliner
  • Inline through blocks
  • Inline methods with rest parameters: ZJIT: Allow inlining methods with rest parameters ruby/ruby#17995
  • Do less in PushInlineFrame ZJIT: Reduce cost of PushInlineFrame #1008
  • Tune it: thresholds, limits, other heuristics
• Support long(er) jumps within a code unit for large blobs of generated code (e.g. templates, filters, protoboeuf)
• Speed up allocation paths: collaborate with GC team to make object allocation faster
• Speed up profiling
• Improve Ractor support: don’t bail if Ractors are enabled
• Support exception handlers ZJIT: Compile exception handlers #966 / throw ZJIT: Compile throw instruction #1016
• ZJIT: Support lifetime holes in the register allocator ruby/ruby#18143
• Continuous perf/benchmark comments mentioned on PRs